Project Composition
Structure of Fiesta project
Last updated
Structure of Fiesta project
Last updated
Fiesta project is an instance of federated learning applied on spectrum data. Just like vanilla federated learning, edge nodes first collect data by themselves and perform machine learning. After training with local data, edge nodes will upload trained model, instead of raw data, to the average server. Average server will gather all received local models and generate a new global model for the next round local training.
In the context of Fiesta project, our edge nodes are actually mobile devices connected with a compatible SDR sampling board. The board will sample spectrum data and send it to android device. With Fiesta Android Client installed, this mobile device can perform federated learning with no extra efforts.
For Fiesta project, we did not use a single average server. As a substitution, we adopt a cluster of miner nodes. Each miner nodes will do the almost the same thing as an average server in terms of model weights integration. However miners will share and record those local models in a blockchain manner, just like distributed ledger.
In this case, the system becomes much more robust to node failures as well as network failures compared with a centralized server. Additionally, transparent and immutable records of each member's contribution makes a fair incentive mechanism practicable, which is the key for maintaining a crowdsourcing ecosystem.
Besides mobile devices and miner nodes, an extra type of node, seed node, is introduced to sustain the connection between participants of Fiesta project, who is pretty similar to tracker server in BitTorrent protocol. It tracks available miners in the system and responses to requests from miners and mobile devices.
Mobile device has no idea about the blockchain among the miner network. It just downloads, trains, and uploads like normal edge node in the federated learning.
One cycle of model training in Fiesta will be as follows:
Seed node injects primitive model and training parameters into miner network
Mobile device fetches the list of those online miners from seed node
Mobile device fetches the latest global model from any miner
Mobile device trains model locally with local dataset
Mobile device sends trained local model to any miner, a miner will also share its received local model with peers.
Miner gathers received local models into one new global model
Besides the normal training cycle, seed node also plays a role in node registration and reward stats.
Seed node fetches the stable chain from miner network and calculate everyone's contribution periodically.
Each Miner needs to register itself to seed node from time to time to make itself popular
Everyone can have easy access to the trained model at any time, which is our goal of crowdsourcing spectrum sensing. Just visit the corresponded API of miners.
